fewer default steps

post.ipynb

@@ -49,7 +49,7 @@
     "strt = time.time()\n",
     "with simple_model:\n",
     "    simple_trace = pm.sample(\n",
-    "        draws=3000, tune=3000, random_seed=42, return_inferencedata=True\n",
+    "        draws=3000, tune=3000, random_seed=42, chains=2, return_inferencedata=True\n",
     "    )\n",
     "\n",
     "    # About half the time is spent in tuning so correct for that\n",
@@ -116,7 +116,7 @@
     "with model:\n",
     "    strt = time.time()\n",
     "    default_trace = pm.sample(\n",
-    "        draws=10000, tune=5000, random_seed=42, return_inferencedata=True\n",
+    "        draws=3000, tune=3000, random_seed=42, chains=2, return_inferencedata=True\n",
     "    )\n",
     "    default_time = 0.5 * (time.time() - strt)\n",
     "\n",
@@ -258,10 +258,19 @@
    "source": [
     "The computational efficiency of this method is similar to PyMC3's default performance on an isotropic Gaussian (within a factor of a few) and corresponds to an improvement of more than *three orders of magnitude* over the default PyMC3 performance on a correlated Gaussian.\n",
     "\n",
-    "While I've found that this procedure can substantially improve the sampling efficiency in many real world scenerios (especially during exploratory phases of a project), you shouldn't forget about reparameterization because that can provide even better performance and help identify problems with your model specification.\n",
-    "Futhermore, this method might run into numerical issues for high dimensional problems because more samples will be needed to reliably estimate the off-diagonal elements of the mass matrix.\n",
-    "Either way, hopefully this is helpful to folks until PyMC3 includes native support for this type of procedure."
+    "While I've found that this procedure can substantially improve the sampling efficiency in many real world scenarios (especially during exploratory phases of a project), you shouldn't forget about reparameterization because that can provide even better performance and help identify problems with your model specification.\n",
+    "Furthermore, this method might run into numerical issues for high dimensional problems because more samples will be needed to reliably estimate the off-diagonal elements of the mass matrix.\n",
+    "Either way, hopefully this is helpful to folks until PyMC3 includes native support for this type of procedure.\n",
+    "\n",
+    "*Edit: This feature is now available in PyMC3 using the* `init=\"adapt_full\"` *argument to* `pm.sample`."
    ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
   }
  ],
  "metadata": {